How many cycles does Iterative Retrieval typically run?

The pattern is designed to run for a maximum of 3 cycles. This provides a balance between finding the most relevant context and maintaining execution speed.

How does this skill improve token usage?

It improves token usage by identifying and excluding irrelevant files in early cycles and only returning files with a high relevance score (typically 0.7 or higher).

What is the 'context problem' in multi-agent workflows?

The context problem occurs when subagents are spawned without knowing exactly which files they need. This leads to either missing critical information or exceeding token limits by sending too much irrelevant data.

Does this skill work with any programming language?

Yes, it uses pattern matching, keyword analysis, and relevance scoring that are language-agnostic and adaptable to any codebase structure.

Iterative Context Retrieval

Name: Iterative Context Retrieval
Author: flatrick

byflatrick

0•

开发者工具

Refines codebase context for AI agents through a multi-phase loop to ensure subagents possess precise information for complex tasks.

The Iterative Retrieval skill implements a sophisticated 4-phase loop—Dispatch, Evaluate, Refine, and Loop—to solve the context discovery problem in multi-agent workflows. By progressively narrowing down relevant files and learning project-specific terminology, it prevents 'context too large' errors and missing information failures. This skill is essential for developers building complex agentic systems where subagents must autonomously explore large codebases with high precision and low token overhead, ensuring that Claude only processes the most relevant code snippets.

主要功能

01Four-phase retrieval loop: Dispatch, Evaluate, Refine, and Loop

02Token usage optimization by aggressively excluding irrelevant file paths

030 GitHub stars

04Dynamic terminology discovery to learn project-specific naming conventions

05Automated relevance scoring (0-1 scale) to filter out noise

06Gap identification to target missing context in subsequent search cycles

使用场景

01Implementing new features that require cross-referencing multiple disparate modules and patterns

02Spawning subagents to fix bugs in large, unfamiliar codebases where dependencies are unclear

03Building RAG-like pipelines for autonomous code exploration and internal documentation

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add flatrick/mdt iterative-retrieval

For use in Claude.ai and ChatGPT